Implementation of End-To-End Memory Networks with sklearn-like interface using Tensorflow. Tasks are from the bAbl dataset.
This is a fork of the original repository https://github.com/domluna/memn2n written by Dominique Luna.
Changes from the original implementation include -
- L2 regularizations for weight matrices
- Jaccard similarity for sentence selection to form memories
- Per task early stopping during joint training
- Web demo (see below)
Task | Testing Accuracy |
---|---|
1 | 99.6 |
2 | 67.4 |
3 | 57.6 |
4 | 98.4 |
5 | 83.1 |
6 | 99.3 |
7 | 85.8 |
8 | 91.8 |
9 | 99.3 |
10 | 95.7 |
11 | 97.5 |
12 | 99.2 |
13 | 98.3 |
14 | 87.7 |
15 | 100 |
16 | 48 |
17 | 61.3 |
18 | 92.1 |
19 | 10.8 |
20 | 100 |
We added a web demo allowing us to test the model and visualize the memory probabilities in each hop (episode). Below is an example that demonstrate it -